Context-aware bioinformatic AI agents

Differential gene expression analysis follows a fixed order: genes are ranked by a statistical test, and biological meaning is attached afterwards through enrichment against ontologies. The interpretation step is where most of the prior knowledge enters, and it enters late, in aggregate form and without direction, so a gene list is rarely sufficient to discriminate between competing mechanistic explanations. Meanwhile the relevant knowledge, including the sign of regulatory relationships and the context in which they hold, is distributed across the literature and curated databases in a form that is not directly computable.

This project inverts the order. Prior knowledge is first extracted from publications and curated resources into a context-specific knowledge graph, structured by an ontology that fixes the schema of the extracted relations and preserves their direction and provenance. The graph then generates explicit, directional predictions of gene regulation by signed propagation from seed nodes, and experimental data are evaluated against those predictions through a coherence score, so that the informative output becomes the disagreement rather than the ranking. The case study is resistance to anti-PD-1 immunotherapy in clear cell renal cell carcinoma. The current phase concerns the extraction pipeline and the reliability of the constructed graph; validation against patient cohort data follows.

Institutions

Collaborators

Funding